llama32-3b-ecra-sft (Earnings Call Research Assistant)

QLoRA adapter on unsloth/Llama-3.2-3B-Instruct for grounded financial research Q&A / summarization over public earnings-call style text. Paired in the GitHub repo with hybrid BM25 + dense retrieval. Metrics below are copied from machine-written JSON only — never hand-invented.

• GitHub: https://github.com/nuwanda94/earnings-call-research-assistant
• RAG report: https://github.com/nuwanda94/earnings-call-research-assistant/blob/main/evals/reports/RAG_EVAL_REPORT.md

Intended use

Research-style questions over retrieved public excerpts (guidance, margins, segment color, named risks). Prefer refusing figures that are not in context. Not investment advice.

Training

Item	Value
Base	unsloth/Llama-3.2-3B-Instruct (4-bit QLoRA)
Seed	3407
Dataset	ecra-sft-v0.1.0
Train / val rows	2127 / 240
Effective batch	16
Adapter path	/kaggle/working/earnings-call-research-assistant/outputs/adapters/llama32-3b-ecra-sft
Card generated (UTC)	2026-09-14T08:35:37Z

Retrieval corpus (when RAG was run)

Item	Value
N chunks	19990
Documents	3
Corpus version	v0.1.0
Embed model	sentence-transformers/all-MiniLM-L6-v2
Dense backend	sentence-transformers
Eval queries	50

Measured metrics (hybrid retrieval)

k	Recall	nDCG
1	TBD	TBD
3	TBD	TBD
5	TBD	TBD
10	TBD	TBD

Grounded generation

Side	citation-hit	token F1	grounded accuracy	dry_run
base	0.5934	TBD	0.5800	False
adapter	0.8760	TBD	1.0000	False

How to load

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "unsloth/Llama-3.2-3B-Instruct"
adapter = "<this-repo-id>"  # e.g. nuwanda94/llama32-3b-ecra-sft
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, device_map="auto")
model = PeftModel.from_pretrained(model, adapter)

Reproduce metrics from the GitHub repo:

python scripts/eval_retrieval.py --run
python scripts/eval_rag_generate.py --run --adapter-dir outputs/adapters/llama32-3b-ecra-sft

Limitations

Public data only. Report measured N and query counts; fixture-scale runs are not SEC coverage. Citation-hit ≠ numerical correctness. Token: HF_TOKEN / huggingface-cli login only — never commit secrets.

Notes from card builder

• (none)